Applying a Deep Convolutional Neural Network to Segment the Flooding Extent from Images of Urban Flooding: Toward an Urban Flood Monitoring System

Cities across the globe are becoming increasingly vulnerable to flooding primarily due to climate change and urbanization. New approaches are needed to monitor these flooding impacts for real-time decision support and longer-term adaption planning. The longer-term goal of this research is to build a camera-based flood monitoring system that automatically extracts flooding extents from camera imagery to enhance emergency response efforts. In comparison to common flood monitoring tools, surveillance cameras have the advantages for flood monitoring in urban environments of high-resolution imagery and wide existing usage in municipal monitoring networks. Toward the longer-term goal of this research, the objective of this study was to use Deeplabv3+, a model based on deep Convolutional Neural Network (CNN), for semantic segmentation of flood images. Two datasets containing 253 and 1039 images of flooding in an urban context, respectively, were split at a ratio of 4:1 for training and validation. The mean Intersection over Union (mIoU), recall, and precision metrics were used to assess the performance of the model for both training and validation. With higher efficiency and fewer parameters, a pretrained Linknet model was used for baseline comparison against the trained Deeplabv3+ model. Results show that both models extracted flood extent from the images at a good accuracy, with Deeplabv3+ performing better. Deeplabv3+ achieved a mean mIoU of 87.5%, a mean recall of 95.6% and a mean precision of 91.2% in all validations, while Linknet achieved 79.2%, 90.4% and 86.8% for those three metrics. The validations on the same dataset were slightly better than those across datasets, which is expected but shows the transferability of the model. The factors that affect the performance of the models were primarily night conditions with poor visibility, high reflectance from strong light sources, and wavy water surfaces due to water flow or vehicles passing through ponded water. Based on these findings, it can be concluded that a CNN-based model like Deeplabv3+ can extract flood extent from camera images of flooding in an urban context. With additional steps to georeference the segmented flood extent, it will be possible to establish a camera-based flood monitoring framework that can report real-time extent and depth of urban flood.